Round Corner
Department of Computer and Information Science


Explainable Anomaly detection for wind farm export cable (AI Lab pitch)

Background: Equinor currently operates several offshore wind parks: Dudgeon, Sheringham Shoal and Hywind Scotland, with more to come in the future. All the power produced is carried to shore via one or more export cables. For Dudgeon wind farm, there are two such cables, and they are fitted with a Distributed Temperature System (DTS) where refracted light through a fibre optic cable is used to determine the temperature along the cable. The temperature varies a lot along the cable and with the amount of power going through it. In addition, things like sea temperature, burial depth and seabed density may affect the cable temperature. Hence, it is not trivial to determine whether a given temperature reading is normal or not. An abnormal temperature reading could warn of an impending failure in what is a critical asset for the wind farm, so we would very much like to pick it up as early as possible!

The DTS data, along with data for electrical current, sea temp., burial depth and seabed plow resistance is made available in Equinor’s OMNIA data platform in Microsoft Azure, where it is stored in a Data lake component (HDFS). The combination of scalable storage and scalable compute power available in Azure opens up new possibilities for how to use these data, one of which is to train machine learning models to detect any anomalous temperature readings in the cables.

We have had a master student working on this topic in 2018/2019, so there will be results that can be taken as a starting point for further work.

Data and compute.
We have data from the DTS system since November 1st 2017. You get the temperature at one-meter intervals along each 40 km cable. The other data types are available for the same time span (where relevant) and. Note that the data does not, as far as we know, contain any examples of previous problems with the cables. The raw data will be available in text file format (CSV, JSON or similar) on our cloud-hosted data lake. A Spark cluster is available for efficient data transformation. Other compute resources are readily available in the Azure cloud platform, including GPU nodes.

The task.

  1. Develop one or more models that can be used to determine if a temperature reading for a segment of the cable is anomalous or not.
    The current master student is looking at this problem using various forms of standard neural networks. We would ideally like for next year’s student(s) to attempt more explainable (e.g. Baysian) methods, so that the predictions are easier to interpret and gain confidence in. Keywords: Probabilistic programming, Tensorflow Probability, PyMC, Edward, Infer.NET
  2. Set up a system for automatic retraining of the models developed.

Task 2) may be implemented by or in collaboration with the Equinor team, if the scope is too big for the student(s).

The student(s) will be offered a desk at Equinor’s research centre at Rotvoll, to be able to work closely with the team responsible for the wind farm data in the OMNIA data platform.

Secondary supervisor from Equinor.

Vidar Slåtten, and potentially others from the same team in Equinor, will act as the secondary supervisor(s) on the thesis.



Other information:

This is a project in collaboration with an external partner. If you choose this project, then I will serve as the responsible from NTNUs side, but the actual work will also be in tight collaboration with personell from the external partner as listed above.

If you consider picking a project with me as the supervisor, then please look at



Helge Langseth Helge Langseth
310 IT-bygget
735 96488 
NTNU logo